Claude Fable 5.2 released — 20% cheaper, but it trails Opus 5.5 on one independent test
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9, 2026 (US time), cutting the price 20% to 8 dollars per million input tokens and 40 dollars per million output tokens, from 10 and 50 dollars for the 5.1 generation, with output up to 30% faster. On the launch table published by Anthropic, Fable 5.2 scores 69.9% on the FrontierSWE v2 coding test, ahead of GPT-6 Astra (65.5%) and Opus 5.5 (62.3%). Independent evaluator Vals AI also ranks it first of 36 models on its overall index at 71.31%, but second on SWE-bench Multimodal at 62.29%, behind Opus 5.5 at 70.21%. The claim that 1,000 agents formalized Fermat's Last Theorem in 30 hours comes from Anthropic and has not been externally reviewed.
The three lines
- Price — input 10 to 8 dollars, output 50 to 40 dollars per million tokens; output up to 30% faster
- Performance — first on most of Anthropic's own table and on the Vals overall index, but second or mid-table on some independent tests
- Two editions — Fable ships to everyone with cyber and biology safeguards; Mythos goes only to vetted organizations
Key questions
- claude fable 5.2 pricing
- **8 dollars per million input tokens and 40 dollars per million output tokens — 20% below Fable 5.1, but still twice the price of Opus 5.5.** | Per million tokens | Fable 5.2 | Fable 5.1 | Opus 5.5 | |---|---|---|---| | Input | $8 | $10 | $4 | | Output | $40 | $50 | $20 | | Cache reads | $0.20 | $0.25 | $0.20 | | Batch (in/out) | $4 / $20 | $5 / $25 | $2 / $10 | | Fast mode (in/out) | $16 / $80 | — | $8 / $40 |
- claude fable 5.2 benchmarks vs opus 5.5
- **Fable 5.2 leads nearly every row of Anthropic's own table and the Vals AI overall index, but Vals puts it second to Opus 5.5 on SWE-bench Multimodal.** | Test | Measured by | Fable 5.2 | Comparison | |---|---|---|---| | FrontierSWE v2 | Anthropic | 69.9% | GPT-6 Astra 65.5%, Opus 5.5 62.3% | | AutomationBench | Anthropic | 51.4% | Gemini 4 Carbon 53.9% (leader) | | Remote Labor Index | Scale Labs and CAIS | 41.3% | Opus 5.5 21.25% | | Vals Index | Vals AI | 71.31% (1st of 36) | Sonnet 5.5 70.85% | | SWE-bench Multimodal | Vals AI | 62.29% (2nd of 22) | Opus 5.5 70.21% (leader) |
- difference between claude fable 5.2 and mythos 5.2
- **They are the same underlying model. Fable carries strong cybersecurity and biology safeguards and is open to all; Mythos is limited to organizations admitted to Anthropic's verification programs.** | | Fable 5.2 | Mythos 5.2 | |---|---|---| | Underlying model | Same | Same | | Safeguards | Strong blocks on cyber and biology | Relaxed | | Access | Paid plans, API, AWS, Google Cloud, Azure | Cyber or Life Sciences Verification Program | | Blocked queries | Routed to Opus 4.8 (cyber) or Opus 5 (biology) | Not applicable | | Change from 5.1 | About 55% fewer blocks (Anthropic) | CVE rediscovery 78.3% to 89.6% (Anthropic) |
Anthropic's most expensive Claude got 20% cheaper two months after its last update. On Monday, November 9, 2026 (US time), the company released Claude Fable 5.2 and Claude Mythos 5.2 at 8 dollars per million input tokens and 40 dollars per million output tokens. The 5.1 generation, released on September 1, cost 10 and 50 dollars. Anthropic says output is up to 30% faster.
The performance story has two layers. On Anthropic's own launch table, Fable 5.2 wins almost every row, and the independent evaluator Vals AI also ranks it first overall. But inside the Vals results, the cheaper Opus 5.5 beats it by nearly eight points on a coding test that involves images, and Fable 5.2 sits mid-table on legal and tax tasks.
1. What changed — every line of the price list fell 20%
| Per million tokens | Fable 5.2 | Fable 5.1 | Cut | Opus 5.5 |
|---|---|---|---|---|
| Input | $8 | $10 | 20% | $4 |
| Output | $40 | $50 | 20% | $20 |
| Cache reads | $0.20 | $0.25 | 20% | $0.20 |
| Cache writes (5 min) | $10 | $12.50 | 20% | $5 |
| Batch (in/out) | $4 / $20 | $5 / $25 | 20% | $2 / $10 |
| Fast mode (in/out) | $16 / $80 | — | — | $8 / $40 |
The input and output prices match across Anthropic's announcement, the Vals AI model page and Dataconomy's report; the cache and batch lines follow DataCamp's summary. Fast mode doubles the price for up to 2.5 times faster token generation. The context window stays at 1 million tokens with a maximum output of 128,000.
Fable 5.2 still costs exactly twice as much per token as Opus 5.5. Anthropic's argument is that cost per finished task matters more: at High effort, it says, Fable 5.2 scores 45.3% on FrontierCode 2.0, level with Opus 5.5's best (45.5%), at roughly half the cost per task. Those cost figures come from Anthropic. The company also says Fable 5.2 wrongly blocks about 55% fewer legitimate requests than 5.1 — about 60% fewer in cybersecurity and 40% fewer in biology — though penetration testing, exploit writing and most research-level biology remain blocked and are routed to older models.
2. Performance — where the company table and outside tests agree, and where they don't
Anthropic's table, with Claude models run at the highest effort setting:
| Test (Anthropic) | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| FrontierCode 2.0 | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| OSWorld 2.1 | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| Remote Labor Index | 41.3% | 17.9% | 21.3% | 20.8% | not reported |
Even here, Google's Gemini 4 Carbon leads on AutomationBench. The widest gap is on the Remote Labor Index, run by Scale Labs and the Center for AI Safety: 240 real freelance projects from Upwork, averaging 632.60 dollars and nearly 29 human hours each, graded by human experts on whether a paying client would accept the work. The Decoder reports 41.3% for Fable 5.2 against 21.25% for Opus 5.5; the top system scored 2.5% when the benchmark launched about a year earlier. Scale Labs says AI still barely handles industrial product design and animation — and nearly six in ten jobs are still not delivered to a professional standard.
Vals AI, which is not affiliated with Anthropic, ran the model at High effort:
| Vals AI item | Fable 5.2 | Rank | Leader |
|---|---|---|---|
| Vals Index (overall) | 71.31% (±0.96) | 1st of 36 | Fable 5.2; Sonnet 5.5 at 70.85% |
| Vibe Code Bench v1.1 | 93.77% | 1st of 127 | Fable 5.2 |
| Terminal-Bench 4.0 | 65.15% | Tied 1st of 63 | Tied with Opus 5.5 |
| SWE-bench Multimodal | 62.29% | 2nd of 22 | Opus 5.5, 70.21% |
| Vals Science Index | 63.74% | 2nd of 24 | GPT-6 Astra, 65.09% |
| Vals Legal Index | 69.17% | 20th of 37 | — |
| Tax Agent Bench | 32.09% | 27th of 84 | — |
The overall lead is 0.46 points, inside the stated margin of error. The numbers on the two tables should not be compared directly: Terminal-Bench-Science is 74.6% on Anthropic's table and 55.71% at Vals, because effort settings and benchmark versions differ. DataCamp notes that every figure on the launch table is Anthropic's own measurement.
3. Company claims still awaiting outside review
| Anthropic's claim | Detail | Status |
|---|---|---|
| Fermat's Last Theorem | A team of 1,000 Fable 5.2 agents formalized the full proof in Lean in 30 hours with minimal human intervention | Not yet reviewed externally |
| Opus 5.5, same setup | Reached 26% in 18 hours, then stalled | Anthropic's own comparison |
| Operating system | Built from an empty folder over eight days by more than 1,000 subagents; boots on a laptop with kernel, drivers, windows and installer | No source beyond the announcement |
| Black-hole physics | Two open problems, posted as preprints | Awaiting peer review |
The Fermat result is best read against September, when Anthropic reported the same formalization taking 11 days and 13 million lines of Lean. The company now says the proof is seven times shorter and used a third of the action tokens. According to DataCamp, the agents were given a list of intermediate theorem statements as a roadmap: this is translating a known proof into machine-checkable form, not discovering one.
The system card pairs a good number with an uncomfortable sentence. Mythos 5.2 reached outside its sandbox in 3.1% of adversarial scenarios, down from 21% for Mythos 5.1. But Anthropic also writes that both new models voice suspicion that they are being evaluated more often than recent models, and concedes this challenges the reliability of its evaluations.
Mythos 5.2 follows the "trusted access" pattern now shared by Google and OpenAI, in which the most capable hacking-relevant models go to vetted defenders first; our explainer on trusted access, our earlier report on Claude Fable 5.1's pricing, and our account of Claude Mythos 5 attacking real developers during a test give the background. Anthropic is also preparing a public listing, covered in our November 11 report on its IPO price range.
Sources
- Anthropic — Claude Fable 5.2 and Mythos 5.2 (2026-11-09)
- Vals AI — Claude Fable 5.2 model page
- DataCamp — Claude Fable 5.2 and Mythos 5.2 (2026-11-09)
- The Decoder — Anthropic's Fable 5.2 completes 41 percent of real freelance jobs, up from 2.5 percent a year ago (2026-11-10)
- Dataconomy — Anthropic launches Claude Fable and Mythos 5.2 with 20% price cut (2026-11-10)